The AI that tidies your writing has a slant of its own

The tools your team uses to tidy up writing do more than fix grammar. A new study from the Oxford Internet Institute finds they also nudge the position of what you write, even when told to keep the meaning.

The tools are not dangerous, and this is not a reason to stop using them. The catch is that the nudge is invisible, and it runs through everything a firm now drafts with AI. For a business that sells judgement in words, that is worth a closer look.

What the study actually found

The researchers took human-written text on contested subjects and asked several leading models, from different makers, to improve it. The edited versions came back leaning in a consistent direction, even when the instruction was simply to preserve the original meaning. Different models tended to lean the same way as each other, which rules out a single vendor's quirk.

This is not a lab curiosity. The paper's own opening notes that these models now polish users' posts on LinkedIn and add context to posts on X. The editing it studied is the editing your team already does every day.

The study then models what happens when this sits between people on a network. Small, repeated nudges do not cancel out. They accumulate, and in the researchers' simulations on real social-network data they can shift the overall opinion of a group in the direction of the slant.

In one case the researchers traced a model's lean back to a single instruction the platform had given it. That shows the slant is partly a design choice, rather than only an accident of training.

Two caveats keep this honest. First, the headline concern in the paper is societal, about public opinion at scale, and not about any one firm. Second, the topics tested were politically contested ones, chosen because a slant is easy to measure there.

The read for your business is an extension of the same mechanism, and it holds up for a simple reason. A model rewrites opinion-bearing text and leaves it saying something slightly different from what you meant.

Why this reaches a professional-services firm

Picture where AI has quietly become part of how writing gets done. A proposal gets reworded to sound crisper. A client update is softened before it goes out.

A candidate rejection is made warmer. A post in the firm's name is improved before it goes up. Each of those is an edit, and each edit is a chance for the position to move a notch.

The work most exposed is the opinion-bearing work, where the value is the view rather than the words. Think of a recommendation in an advice note, a firm position in a report, or the point of view your marketing is known for.

In a neutral status email a small shift in tone costs little. In a recommendation, the stance is the product, so a shift in stance changes what the client is actually buying.

Here is how small the move can be. An adviser writes that the numbers suggest caution, and the tidy-up returns that the numbers point to holding off. The second version is firmer than the file supports.

Nobody chose to harden that view. The tool did, and it now reads as the firm's opinion.

Keep the people on the other side in view too. A candidate reads the tone of a rejection and decides whether to apply again. A client reads the confidence of an advice note and acts on it. If the model makes the note breezier, or the rejection warmer, than you intended, the person reading it takes away a position you did not choose to send.

How this differs from the problems you already know

This is not the accuracy problem. That one is about the model getting a fact wrong, and the fix is a human check before the work reaches a client. A slanted edit can be perfectly accurate and still off, because nothing is wrong and the emphasis has simply moved.

It is also not the same as a model changing under you over months. That slow, silent behaviour drift is real and worth watching, but it runs on a different clock. The slant here is present on the very first edit, today, in a model behaving exactly as designed.

It helps to name it plainly. This is a slant on every edit, a standing feature of the tool rather than an occasional fault. Once you see it that way, the response gets simpler.

What to do about it

The answer is not to pull AI out of your writing, and it is not to turn every email into a review meeting. Three habits cover most of the risk.

1. Decide the position before you open the tool. Write the line you actually mean first, in your own words, even roughly. Then let AI tighten it. If you draft from a blank prompt inside the tool, you have nothing to compare against, so you cannot tell what it moved.

2. On anything that carries a view, read the edit for stance, not only for errors. The question is quick. Does this still say what you meant, and take the position you take? That read is different from proofreading, and it is the one that catches a moved position.

3. Keep one human voice on the pieces that carry the firm's view. Think of the marketing you are known for, a signed recommendation, or a public post in the firm's name. These are where a consistent slant compounds over months into a voice that is not quite yours.

This is the ordinary operating judgement our AI coaching for leaders is built around. The useful question is how to keep your firm's voice and view intact while AI does the tidying, not which tool to buy.

So the decision this week is small and specific. Pick the two or three things your firm publishes that carry its view, and add one question to the final read: does this still take our position, or the model's?

If it would help to find where the slant is creeping into your own writing, book a discovery call. We will look at the work you actually send to clients and set a check that fits how your team works.

Does using AI to edit our writing change the meaning?

It can, in a specific way. The Oxford research found that leading models, asked to improve opinion writing, tended to shift its position in a consistent direction, even when told to preserve the meaning. On plain factual or administrative text the effect is minor. On writing that carries a view, say a recommendation, a firm position or a marketing line, it is worth a deliberate check.

How is this different from the AI just making mistakes?

A mistake is the model getting something wrong, and you catch it by verifying facts. A slant is different, because the output can be fluent, accurate and still take a position you did not intend. The check is not "is this correct" but "does this still say what we meant". Both matter, and they are separate reads.

What is the simplest safeguard for a small team?

Decide the position yourself before you use AI to polish it, then read the result once for stance rather than only for typos. For anything published in the firm's name, keep a named person who owns the final voice. That costs about a minute per piece and catches the drift while it is still small.